OpenAI Astra launch imminent as GPT-6 codename leaks, 52% odds

OpenAI's Astra, which solved 10 decade-old math problems for roughly $2,000 in token costs, is set for imminent release, with a 52 percent Polymarket probability pointing to an Aug. 20 launch.

"Astra is very powerful, and we will prepare safety and then make it available to everyone," Sam Altman, chief executive of OpenAI, said in a post responding to the model's cybersecurity assessment.

The model, reportedly built on roughly 10 trillion parameters using a mixture-of-experts architecture under the pretraining project "Doug," cleared OpenAI's "Critical" threshold in the Preparedness Framework's cybersecurity evaluation on Aug. 7. That assessment found Astra could independently identify and develop zero-day exploits across multiple severity levels. The same day, references to the internal codename "mewfour" were scrubbed from 52 pull requests in OpenAI's public Codex repository — two hours and 34 minutes after the security disclosure — only to resurface in the repo's commit history six days later.

The release would mark the first frontier model trained natively for multi-agent collaboration from pretraining, a structural shift that could accelerate enterprise AI adoption and pressure rivals Anthropic and Google in coding and agentic workloads.

The Mewfour Trail

The codename leak traces a concrete path through OpenAI's public development infrastructure. On Aug. 1, the company published a 249-page mathematics paper claiming an internal Astra version resolved 10 open problems in high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography and extremal combinatorics — including the first explicit construction of a non-sofic group, a problem mathematicians pursued for decades.

At current GPT-5.6 Sol API rates — $5 per million input tokens and $30 per million output tokens — generating all 10 proofs would cost approximately $2,000. That figure, roughly the price of a mid-range laptop, reframes a decade of mathematical output as a single API call.

The Aug. 7 security disclosure added a second layer. OpenAI's Preparedness Framework classifies cybersecurity capability into tiers, with "Critical" as the highest. At that level, a model can independently identify and develop zero-day vulnerabilities across multiple severity levels or execute end-to-end attack strategies against hardened targets.

The timing of the "mewfour" cleanup — 52 pull requests scrubbed within hours of the security post — suggests OpenAI was managing internal terminology as the model moved closer to production. The subsequent appearance of the same identifier in the openai/codex repository, where commits show the model being tested at "xhigh" reasoning levels, points to release-candidate status.

Multi-Agent Native Architecture

The more consequential story may be architectural. The Information reported that Astra was trained from pretraining to enable multiple agents to collaborate over extended periods on complex problems. If confirmed, this would be the first frontier model with end-to-end multi-agent coordination baked in at the training stage, rather than bolted on afterward.

OpenAI's existing GPT-5.6 family — Sol (flagship), Terra (value) and Luna (fastest, cheapest) — already layers multi-agent infrastructure on top of separate models. The v2 system introduced inter-agent communication between Sol and Terra during parallel execution, and Codex recently added Luna as a sub-agent. But Dan McAteer, a technology analyst, noted that native multi-agent pretraining would give Astra inherent long-horizon collaboration capabilities out of the box, allowing it to delegate lighter tasks to cheaper models like Luna in Codex.

That capability gap is what current architectures cannot patch with software updates, McAteer said. It mirrors how Claude Code's agentic coding approach forced a fundamental rethink of the underlying logic layer.

The competitive stakes are measurable. Anthropic's Claude Fable currently leads in several coding benchmarks, and Google's Gemini 3.5 Pro has faced internal delays against performance goals, with seven consecutive weekly Kalshi release contracts settling No between June 30 and Aug. 7. An Astra launch this week would reset the frontier — and the pricing power that comes with it.

For investors, the release carries direct implications for Microsoft, which holds a significant stake in OpenAI and integrates its models into Azure; Nvidia, whose GPUs power the training runs; and the broader AI infrastructure complex that has poured nearly $600 billion into capital expenditures this year. A successful Astra launch could accelerate enterprise adoption and justify continued AI infrastructure spending, while a delay would leave the sector's valuation narrative hanging on a model that has already been graded "Critical" for cyber risk.

This article is for informational purposes only and does not constitute investment advice.